Provider Catalog

Canonical source: docs/07-integrations/provider-catalog.md — this page is rendered from it on every publish; edit the source.

ClawAI AI Provider Integration Catalog

Runtime-managed OpenAI-compatible providers (ADR-157)

Admins with ADMIN_CONNECTORS_MANAGE can register a custom compatible provider and toggle built-in provider availability in Connectors > Providers. Built-ins are status-only: compatible runtime settings remain in CONNECTOR_PRESETS, and bespoke providers keep their current adapters. Existing connector rows link to their built-in definitions without rewriting credentials or model history. Custom definitions use the generic adapter for OpenAI chat completions and OpenAI-shaped model listing. NVIDIA NIM's hosted service uses https://integrate.api.nvidia.com; a mocked /v1/models response is covered by the compatibility acceptance test. No live provider call or free-tier claim is made by this feature. Hugging Face task APIs, Pollinations media, and AI Horde's asynchronous protocol need separate adapter implementations.

This document catalogs all AI providers integrated with ClawAI, including setup instructions, available models, capability flags, and classification metadata.

---

Provider Overview

Provider	Type	Models	Streaming	Tools	Vision	Audio	Privacy Class	Cost Class	Latency Class
OpenAI	Cloud	GPT-4o, GPT-4o-mini	Yes	Yes	Yes	No	CLOUD	STANDARD	LOW
Anthropic	Cloud	Claude Sonnet 4, Claude Opus 4	Yes	Yes	Yes	No	CLOUD	STANDARD-HIGH	MEDIUM
Google Gemini	Cloud	Gemini 2.5 Flash	Yes	Yes	Yes	No	CLOUD	LOW	LOW
DeepSeek	Cloud	deepseek-chat	Yes	Yes	No	No	CLOUD	LOW	MEDIUM
Local Ollama	Local	5 models (see below)	No	No	No	No	LOCAL	FREE	VARIABLE

---

1. OpenAI

Overview

OpenAI provides GPT-series large language models. ClawAI integrates via the OpenAI Chat Completions API, which is also the de facto standard for many compatible providers.

Endpoint

• Base URL: https://api.openai.com/v1
• Chat Completions: POST /chat/completions
• Authentication: Bearer token in Authorization header

Setup Instructions

1. Obtain an API key from platform.openai.com.
2. In ClawAI, navigate to Connectors > Add Connector.
3. Select provider: OPENAI.
4. Enter the API key. It will be encrypted with AES-256-GCM before storage.
5. Leave the base URL empty (defaults to https://api.openai.com/v1).
6. Click "Test Connection" to verify.
7. Click "Sync Models" to populate available models.

Models

Model	Context Window	Best For	Streaming	Tools	Vision	Audio
gpt-4o	128K tokens	General purpose, creative writing, complex reasoning	Yes	Yes	Yes	No
gpt-4o-mini	128K tokens	Fast responses, simple tasks, cost optimization	Yes	Yes	Yes	No

Routing Rules (AUTO Mode)

Task Type	Routes To
Creative writing, general chat	gpt-4o-mini
Complex multi-step reasoning	gpt-4o

Cost Class: STANDARD

• gpt-4o: Input $2.50/1M tokens, Output $10.00/1M tokens (approximate)
• gpt-4o-mini: Input $0.15/1M tokens, Output $0.60/1M tokens (approximate)

Latency Class: LOW

• gpt-4o-mini: Typical 0.5-2s for short responses
• gpt-4o: Typical 2-8s for complex queries

Privacy Class: CLOUD

• Data is sent to OpenAI's servers.
• Subject to OpenAI's data usage policies.
• Not recommended for highly sensitive data; use LOCAL_ONLY or PRIVACY_FIRST routing modes instead.

Notes

• OpenAI's Chat Completions API format is used as the standard interface. Other cloud providers (Anthropic, Gemini, DeepSeek) are called through OpenAI-compatible endpoints or adapters.
• Rate limits depend on the API tier. ClawAI applies its own rate limiting (default: 100 req/min) on top of provider limits.

---

2. Anthropic

Overview

Anthropic provides the Claude family of models, known for strong code generation, reasoning, and safety. ClawAI integrates via Anthropic's API using the OpenAI-compatible chat completions format.

Endpoint

• Base URL: https://api.anthropic.com/v1
• Chat Completions: POST /chat/completions
• Authentication: Bearer token (API key) in Authorization header

Setup Instructions

1. Obtain an API key from console.anthropic.com.
2. In ClawAI, navigate to Connectors > Add Connector.
3. Select provider: ANTHROPIC.
4. Enter the API key (AES-256-GCM encrypted at rest).
5. Leave the base URL empty (defaults to https://api.anthropic.com/v1).
6. Click "Test Connection" to verify.
7. Click "Sync Models" to populate available models.

Models

Model	Context Window	Best For	Streaming	Tools	Vision	Audio
claude-sonnet-4	200K tokens	Coding, debugging, code review, detailed analysis	Yes	Yes	Yes	No
claude-opus-4	200K tokens	Deep reasoning, architecture decisions, complex problem solving	Yes	Yes	Yes	No

Routing Rules (AUTO Mode)

Task Type	Routes To
Coding, debugging, code review	claude-sonnet-4
Deep reasoning, architecture discussions	claude-opus-4

Cost Class: STANDARD to HIGH

• claude-sonnet-4: Input $3.00/1M tokens, Output $15.00/1M tokens (approximate)
• claude-opus-4: Input $15.00/1M tokens, Output $75.00/1M tokens (approximate)

Latency Class: MEDIUM

• claude-sonnet-4: Typical 2-6s for code-related queries
• claude-opus-4: Typical 5-15s for complex reasoning

Privacy Class: CLOUD

• Data is sent to Anthropic's servers.
• Anthropic has strong data handling policies (no training on API data by default).
• Preferred fallback provider in PRIVACY_FIRST mode (after local models).

Notes

• Anthropic is the preferred provider for coding tasks and deep reasoning in AUTO mode.
• Claude models excel at following complex instructions and generating well-structured code.
• The PRIVACY_FIRST routing mode uses Anthropic as the cloud fallback due to its privacy-favorable data policies.

---

3. Google Gemini

Overview

Google's Gemini models are multimodal, excelling at image understanding, video analysis, web search integration, and file/data analysis. ClawAI integrates via the Gemini API using the OpenAI-compatible chat completions format.

Endpoint

• Base URL: https://generativelanguage.googleapis.com/v1beta
• Chat Completions: POST /chat/completions (via OpenAI-compatible endpoint)
• Image Generation: POST /models/{model}:generateContent
• Authentication: API key as Bearer token or query parameter

Setup Instructions

1. Obtain an API key from aistudio.google.com.
2. In ClawAI, navigate to Connectors > Add Connector.
3. Select provider: GEMINI.
4. Enter the API key (AES-256-GCM encrypted at rest).
5. Leave the base URL empty (defaults to the standard Gemini endpoint).
6. Click "Test Connection" to verify.
7. Click "Sync Models" to populate available models.

Models

Model	Context Window	Best For	Streaming	Tools	Vision	Audio
gemini-2.5-flash	1M tokens	File/data analysis, image understanding, web search, fast general purpose	Yes	Yes	Yes	No

Routing Rules (AUTO Mode)

Task Type	Routes To
Image/video/YouTube/web search	gemini-2.5-flash
File/data analysis	gemini-2.5-flash

Cost Class: LOW

• gemini-2.5-flash: Input $0.15/1M tokens, Output $0.60/1M tokens (approximate, generous free tier)

Latency Class: LOW

• gemini-2.5-flash: Typical 0.5-3s, very fast for its capability level

Privacy Class: CLOUD

• Data is sent to Google's servers.
• Subject to Google Cloud's data handling terms.

Notes

• Gemini is the preferred provider for multimodal tasks (image understanding, video, web search) in AUTO mode.
• The large context window (1M tokens) makes it ideal for file analysis with large documents.
• Also used for image generation via the Imagen integration.

---

4. DeepSeek

Overview

DeepSeek provides cost-effective models specializing in mathematical reasoning, algorithms, and general-purpose chat. ClawAI integrates via DeepSeek's OpenAI-compatible API.

Endpoint

• Base URL: https://api.deepseek.com/v1
• Chat Completions: POST /chat/completions
• Authentication: Bearer token in Authorization header

Setup Instructions

1. Obtain an API key from platform.deepseek.com.
2. In ClawAI, navigate to Connectors > Add Connector.
3. Select provider: DEEPSEEK.
4. Enter the API key (AES-256-GCM encrypted at rest).
5. Leave the base URL empty (defaults to https://api.deepseek.com/v1).
6. Click "Test Connection" to verify.
7. Click "Sync Models" to populate available models.

Models

Model	Context Window	Best For	Streaming	Tools	Vision	Audio
deepseek-chat	128K tokens	Math, algorithms, cost-effective general purpose	Yes	Yes	No	No

Routing Rules (AUTO Mode)

Task Type	Routes To
Math, algorithms, numerical analysis	deepseek-chat (or local phi3:mini)

Cost Class: LOW

• deepseek-chat: Input $0.14/1M tokens, Output $0.28/1M tokens (approximate, among the cheapest cloud options)

Latency Class: MEDIUM

• deepseek-chat: Typical 2-5s depending on complexity

Privacy Class: CLOUD

• Data is sent to DeepSeek's servers.
• Review DeepSeek's data handling policies for your compliance requirements.

Notes

• DeepSeek is the primary choice for mathematical and algorithmic tasks in AUTO mode.
• Extremely cost-effective, making it a good COST_SAVER fallback.
• Vision/image capabilities are not supported.

---

5. Local Ollama

Overview

Ollama is a local AI runtime that runs models on the user's hardware. No data leaves the local machine, providing maximum privacy. ClawAI installs and manages 5 local models automatically on first startup.

Endpoint

• Base URL: Configured via OLLAMA_BASE_URL environment variable (default: http://ollama:11434)
• Generation: POST /api/generate
• Model List: GET /api/tags
• Authentication: None required (local service)

Setup Instructions

1. Ollama is included in the Docker Compose configuration and starts automatically.
2. On first startup, the claw-ollama-service auto-pulls 5 default models.
3. Models are synced to the database and assigned default roles.
4. No API key or external account is required.
5. Additional models can be pulled via the Models page or POST /api/v1/ollama/pull.

Models

Model	Parameters	Size	Best For	Role Assignment	Streaming	Tools	Vision	Audio
gemma3:4b	4B	3.3GB	Default local chat, routing decisions, memory extraction	ROUTER, FALLBACK_CHAT	No	No	No	No
llama3.2:3b	3B	2.0GB	Local reasoning tasks	REASONING	No	No	No	No
phi3:mini	3.8B	2.2GB	Coding, math, structured output	CODING	No	No	No	No
gemma2:2b	2B	1.6GB	Fast general-purpose responses	FALLBACK_CHAT	No	No	No	No
tinyllama	1.1B	637MB	Ultra-fast routing fallback, simple tasks only	ROUTER (fallback)	No	No	No	No

Model Roles

Role	Description	Default Model
ROUTER	Analyzes message content to determine optimal provider/model in AUTO mode	gemma3:4b
FALLBACK_CHAT	Handles chat requests when cloud providers are unavailable	gemma3:4b
REASONING	Handles reasoning and analytical tasks locally	llama3.2:3b
CODING	Handles coding and math tasks locally	phi3:mini

Routing Rules

Routing Mode	Behavior
LOCAL_ONLY	Always uses local-ollama/gemma3:4b. No data leaves the machine.
PRIVACY_FIRST	Uses local models when Ollama is healthy. Falls back to Anthropic only if local is unavailable.
COST_SAVER	Uses local models when healthy. Falls back to cheapest cloud provider.
AUTO	The router model (gemma3:4b) analyzes each message. Routes simple Q&A, translations, and privacy-sensitive content to local models.

Cost Class: FREE

• No per-token costs.
• Only hardware costs (CPU/GPU, RAM, disk for model storage).
• Total disk requirement for all 5 models: approximately 9.7GB.

Latency Class: VARIABLE

• Depends heavily on hardware (CPU vs GPU, RAM availability).
• gemma3:4b: 2-10s on CPU, 0.5-3s on GPU.
• tinyllama: 0.5-2s on CPU (smallest model).
• phi3:mini: 2-8s on CPU for coding tasks.

Privacy Class: LOCAL

• No data leaves the local machine.
• Models run entirely within the Docker environment.
• Ideal for confidential, sensitive, or regulated data.
• The highest privacy guarantee available in ClawAI.

Configuration

Environment Variable	Description	Default
OLLAMA_BASE_URL	URL of the Ollama runtime	http://ollama:11434
OLLAMA_ROUTER_MODEL	Model used for AUTO routing decisions	gemma3:4b
OLLAMA_ROUTER_TIMEOUT_MS	Timeout for routing decisions before heuristic fallback	10000 (10s)
MEMORY_EXTRACTION_MODEL	Model used for automatic memory extraction	gemma3:4b

Health Check

• GET /api/v1/ollama/health?runtime=OLLAMA returns the runtime health status.
• Health is checked by the health aggregator and reported on the Observability dashboard.
• If Ollama is unhealthy, PRIVACY_FIRST and COST_SAVER modes fall back to cloud providers.

Notes

• Models are auto-synced from the Ollama runtime to the database on ollama-service startup.
• Additional models can be pulled from the Ollama registry at any time.
• Model role assignments can be changed via POST /api/v1/ollama/assign-role.
• The router model runs with temperature=0 for deterministic routing decisions.
• Router output is validated with a Zod schema to ensure structured, parseable routing decisions.
• When pulling models, the system sends the full model name with tag (e.g., "gemma3:4b") to the Ollama runtime.

---

Provider Comparison Matrix

Capability Comparison

Capability	OpenAI	Anthropic	Gemini	DeepSeek	Local Ollama
Text chat	Excellent	Excellent	Excellent	Good	Good (limited by model size)
Code generation	Good	Excellent	Good	Good	Fair (phi3:mini best)
Reasoning	Good	Excellent (Opus)	Good	Good (math)	Fair (llama3.2:3b)
Data analysis	Good	Good	Excellent	Good	Poor
Image understanding	Yes (GPT-4o)	Yes (Claude)	Yes (Gemini)	No	No
Image generation	Via DALL-E	No	Via Imagen	No	Via Stable Diffusion (if configured)
Web/search	No	No	Yes	No	No
Privacy	Cloud	Cloud (strong policy)	Cloud	Cloud	Full local
Offline capable	No	No	No	No	Yes

Task-to-Provider Routing (AUTO Mode Summary)

Task Category	Primary Provider	Fallback
Coding, debugging, code review	Anthropic / claude-sonnet-4	OpenAI / gpt-4o
Deep reasoning, architecture	Anthropic / claude-opus-4	OpenAI / gpt-4o
Image/video/web search	Gemini / gemini-2.5-flash	OpenAI / gpt-4o
Math, algorithms	DeepSeek / deepseek-chat	Local / phi3:mini
Creative writing, chat	OpenAI / gpt-4o-mini	Local / gemma3:4b
Simple Q&A, translations	Local / gemma3:4b	OpenAI / gpt-4o-mini
File/data analysis	Gemini / gemini-2.5-flash	Anthropic / claude-sonnet-4
Privacy-sensitive	Local / gemma3:4b	Never cloud

Cost Comparison (Approximate, per 1M tokens)

Provider / Model	Input Cost	Output Cost	Relative Cost
Local Ollama (all models)	$0.00	$0.00	Free
DeepSeek / deepseek-chat	$0.14	$0.28	Very Low
Gemini / gemini-2.5-flash	$0.15	$0.60	Very Low
OpenAI / gpt-4o-mini	$0.15	$0.60	Very Low
OpenAI / gpt-4o	$2.50	$10.00	Medium
Anthropic / claude-sonnet-4	$3.00	$15.00	Medium-High
Anthropic / claude-opus-4	$15.00	$75.00	High

Custom provider recipes (verified 2026-10-04 on the local stack)

Create under Connector providers; every value is data, no code. Keys are stored as UPPER_SNAKE. Chat path is relative to the base URL.

Provider	Base URL	Models path / format	Chat path	Key header	Result
Pollinations AI	https://gen.pollinations.ai	/v1/models, OPENAI_LIST	/v1/chat/completions	Authorization Bearer	96 chat models, 4 of 4 sampled chats answered
Hugging Face	https://router.huggingface.co	/v1/models, OPENAI_LIST	/v1/chat/completions	Authorization Bearer	134 models, 4 of 4 sampled chats answered
NVIDIA NIM	https://integrate.api.nvidia.com	/v1/models, OPENAI_LIST	/v1/chat/completions	Authorization Bearer	70 listed, 1 of 8 sampled callable (key-scoped, see guide)
AI Horde	https://oai.aihorde.net	/v1/models, OPENAI_LIST	/v1/chat/completions	Authorization Bearer	healthy, 27 models; chats over 512 tokens need kudos

AI Horde's native API (https://aihorde.net/api, /v2/status/models?type=text, apikey header) syncs models but has no /chat/completions; use the OpenAI gateway above for chat. type=text is the whole live text set (28); there is no single route for text and image together, and image models cannot chat.